Thursday, September 10, 2026 - 5 Minute AI News

0:00A researcher at Anthropic just publicly resigned, stating a belief that self-improving AI could eventually cause human extinction. We will look at the growing internal tensions at Frontier Labs. Welcome to 5 Minute AI News for Thursday, September 10th, 2026. We also have a new system for mapping human DNA and a smartwatch feature that is always listening to your conversations. Let us start with that resignation at Anthropic. Jacob Coxon, a researcher on the company's AI pre-training team, has left his position. He released a public statement accusing major AI laboratories of gambling with human lives. His departure centers on the earnest belief that self-improving artificial intelligence systems could eventually lead to human extinction. It is a notable development because it highlights the internal fracturing we are seeing within these companies. There is a clear tension between rapid commercial deployment and the existential fears held by some of the researchers actually building the systems. Coxon has a strong background in this space. Having worked as a researcher at OpenAI before moving to Anthropic, the internal conversations

1:03seem quite serious, reflecting a broader anxiety across the industry. Following his departure, another anthropic colleague, Evan Hubbinger, publicly estimated that the odds of a misaligned super intelligent AI wiping out humanity within the next decade are greater than 10%. That 10% figure is sobering when it comes from inside the facility. It certainly reframes the daily push for more capable models. Over in the research sector, Google DeepMind has launched the Alpha Genome Atlas. This is a system that evaluates and predicts the effects of every possible single letter change in the human genome. It is a significant development for genomics and personalized medicine. Most genetic mutations are harmless, but some cause severe disease. By preemptively mapping these impacts, researchers now have an unprecedented tool for diagnosing disorders. The system has already demonstrated clinical utility by successfully identifying a previously overlooked genetic variant as the likely cause of a patient's epilepsy. The scale of the data is remarkable. The system maps approximately 9 billion possible single-base

2:04DNA changes. The resulting data set spans one petabyte, making it more than 30 times larger than DeepMind's landmark alpha-fold database, which previously advanced our understanding of protein folding. Giving clinicians that kind of reference material could fundamentally alter how genetic diseases are diagnosed in the near future. Shifting to the business side, the startup Harvey has raised an additional $550 million in funding. They build generative AI tools specifically for law firms and legal teams. This latest round brings the company's valuation to $15.5 billion. Up from $11 billion just a few months ago, the capital injection indicates that vertical specific applications are continuing to command premium valuations from venture capitalists, even as the broader market cools down. The legal industry is heavily reliant on complex document processing and precedent analysis, which makes it an ideal fit for these tools. The funding round was led by diffusion and light speed venture partners, with participation from Sequoia, Cliner Perkins, and Goldman Sachs. Securing capital from that caliber of investors shows strong confidence in high margins specialized

3:05enterprise software. In the generative audio space, the music platform Suno just released its version 6 model. What stands out here is that it was built in partnership with major music industry players, including Warner Music Group, BMG, and Belief. Suno is shutting down all of its older models, and states this new version was trained entirely from the ground up on a newly licensed data set. This release highlights a clear divide in the music industry. Labels like Warner are choosing to partner with these platforms, while others like Universal and Sony are actively suing them for copyright infringement. The version 6 update introduces multimodal generation capabilities, meaning it accepts inputs from text, audio, and images. It also allows users to partially edit existing songs using text commands. Interestingly, despite confirming the license partnerships, Suno declined to specify exactly which catalogs were included in the training data. Finally, Apple held its fall hardware event, and announced new audio intelligence features for the Apple Watch Series 12 and Watch Ultra 4. The standout feature is called Siri Recap. Siri Recap utilizes the device's microphone to continuously listen to ambient conversations throughout the day, and generate automated summaries

4:09of the user's interactions. Apple insists that raw audio is not saved, but deploying a mainstream always-on recording device certainly normalizes ambient computing. It is already sparking widespread debate around privacy, consent, and how people might alter their behavior when they know they could be recorded by a nearby smartwatch. These features are part of Apple's new health sensing system. The event also featured a new Apple reference image tool to cryptographically prove photos were not AI-generated. That 10% extinction estimate from anthropic researchers will definitely keep the safety debate going. We will continue to track it. Please remember to like, subscribe, and leave a review for the show on Spotify and Apple podcasts. It really helps us out. Thanks for listening and have a great day.